Research On Autonomous Obstacle Avoidance Of Mobile Robot Based On Deep Reinforcement Learning

Autonomous obstacle avoidance is a key technology that enables mobile robots to be widely used in fields such as self-driving cars and delivery robots.Classical obstacle avoidance algorithms may fail to work in unknown environments due to inappropriate parameter settings,which limits the portability and adaptability of the algorithms.The deep reinforcement learning(DRL)methods have strong adaptability by allowing the robots to continuously try and automatically adjust its policy in the simulation environment.However,DRL methods also face problems such as slow training due to sparse rewards and random exploration during the training process.Therefore,this thesis first improves the performance of the classical obstacle avoidance algorithm by introducing safety gap detection,and then combines it with the DRL algorithm to improve the network training efficiency.Finally,by considering the safety of pedestrians in the same workspace as the mobile robot,the robots’ ability to avoid obstacles is improved under human interference.The specific research includes:(1)We propose a mobile robots autonomous obstacle avoidance method based on safety gap detection.Firstly,a safety gap detection method is proposed based on the existing CG gap algorithm.The method uses the position relationship between the robots and the obstacle to calculate the nearest safety gap to the target,avoiding the problem of impassability of the CG gap.In addition,the corresponding gap sub-goal is calculated based on this safety gap and combined with a classical controller to guide the robot toward the collision-free region.Simulation experiments demonstrate that the introduction of the gap sub-goal can improve the obstacle avoidance performance of the classical controller.(2)Based on the above research,we propose a DRL obstacle avoidance algorithm based on the priori information guidance,which includes the gap sub-goal,classical controller,and toggle switch.We first design a sub-goal selection strategy that integrates robot orientation,gap width,and target distance,which calculates the gap subgoal considering robot motion smoothness from the candidate safety gap.Using it as a temporary goal during robot obstacle avoidance alleviates the problem of poor training reward value in dense scenes.In addition,we design a toggle switch,which can switch to the classical controller to adjust the robot behavior in real-time when the DRL controller strategy performs poorly,avoiding the problem of slow training due to random exploration.Experimental results show that the algorithm can improve the training process of the network and the trained SAC-PGD controller outperforms the standard SAC method.(3)Based on the above priori information guidance,we further propose a DRL obstacle avoidance method considering pedestrian safety in order to take into account the safety and comfort requirements of mobile robots in human environments.On the one hand,we further utilize the relative motion relationship between the robots and the pedestrian in the above reward function to penalize the robot’s action when it is close to the pedestrian.On the other hand,we introduce a social attention network into the original critic network structure to enable the critic network to consider the influence of pedestrian factors additionally.Experimental results show that the proposed method enables the robots to avoid humans in advance and achieve more safe movement.